Prosit-XL

Crosslinking mass spectrometry (XL-MS) is one of the few experimental routes to distance restraints inside intact protein complexes, and therefore to protein-protein interaction maps and structural models of assemblies that resist crystallography. Its bottleneck is identification. A cross-linked peptide spectrum contains fragments from two peptides at once, the two rarely fragment equally, and the search space grows quadratically with the database, so false discovery rate control is both harder and more consequential than for linear peptides. Rescoring with predicted fragment intensities has transformed linear peptide identification, but the predictors that transformation relies on were never built for cross-linked species.

Prosit-XL, developed by Mathias Wilhelm's group at the Technical University of Munich with Juri Rappsilber's laboratory at the Technical University of Berlin, extends the Prosit family to exactly that gap. It predicts MS2 fragment-ion intensities for cross-linked peptides across the two chemistries that dominate the field: cleavable cross-linkers (DSSO, DSBU) and non-cleavable ones (DSS, BS3). It is a sibling of, not a successor to, Prosit-PTM, which addresses modified linear peptides from the same lab; the two share the Prosit lineage and the Koina serving layer but nothing else.

The model was posted in December 2024 and published in Nature Communications in 2025, together with a rescoring pipeline that keeps the false discovery rate honest when intensity features are added to the mix.

#Key Features

• Both cross-linker classes in one family: Prosit-XL-CMS2 handles cleavable DSSO and DSBU spectra, Prosit-XL-NMS2 the non-cleavable DSS and BS3 spectra, and Prosit-XL-CMS3 the MS3 spectra of cleavable links.
• Partner-aware prediction: A second sequence encoder takes the linked partner peptide, because the presence and identity of peptide B measurably shapes how peptide A fragments.
• Symmetric data augmentation: Every spectrum is used twice with the two peptides swapped, which doubles the effective training set and yields predictions for both sides of a link.
• Collision-energy calibration: Inherited from Prosit, normalized collision energy is a model input, so an unseen instrument is handled by calibrating the energy rather than by fine-tuning the network.
• FDR-safe rescoring: Percolator is run at the peptide-spectrum-match level rather than the cross-link level, and the minimum of the two PSM scores stands in for match quality, which separates correct from incorrect matches without inflating the reported FDR.

#Technical Details

Each of the two sequence encoders is an embedding layer followed by two bidirectional GRU layers of 512 units; a third encoder maps precursor charge and normalized collision energy through a dense layer, and its output multiplies the combined latent representation elementwise. Two GRU decoders emit the annotated spectrum — a 348-dimensional vector for CMS2 spectra and 174 dimensions for NMS2, covering three charge states and 29 fragment positions. Training used public XL-MS data harvested from PRIDE: 125,727 CMS2, 70,320 NMS2, and 37,938 CMS3 high-quality spectra, representing roughly 31,000, 17,000, and 9,000 unique peptide pairs. That is around 2% of the volume behind the HCD Prosit 2020 model, which is why each Prosit-XL variant was initialized from Prosit weights and transfer-learned rather than trained from scratch.

On external datasets never seen in training, Prosit-XL-CMS2 achieved a median spectral angle of 0.82 and a Pearson correlation of 0.95 for DSSO-linked peptides, with 0.77 and 0.89 for DSS/BS3 under NMS2. Against ground-truth synthetic benchmarks the rescoring pipeline increased confidently identified cross-linked spectrum matches by 14% and peptide pairs by 12% at 1% FDR on a synthetic peptide dataset, and by 34.9%, 33.4%, and 42.7% for spectrum matches, peptide pairs, and protein-protein interactions respectively on a synthetic protein dataset. Applied to a proteome-scale E. coli and M. pneumoniae experiment it delivered up to a 3.4-fold increase in interaction discovery, and it deepened a spatially resolved interactome map of intact human cytomegalovirus virions enough to surface previously unobserved host-virus contacts.

#Applications

Structural biologists using XL-MS to model complexes, and interactomics groups mapping which proteins touch which, are the direct users. The models are served through Koina and consumed by the Oktoberfest rescoring pipeline, so a laboratory adds them to an existing xiSEARCH or Scout workflow without local GPU training. The virion study illustrates the payoff: more restraints per experiment translates directly into more contacts resolved in a crowded native sample.

#Impact

Prosit-XL closes the gap between linear and cross-linked peptide identification, bringing the intensity-based rescoring that has become routine in shotgun proteomics to a technique where the scoring problem was harder and the tooling thinner. The accompanying FDR analysis is arguably as consequential as the predictor: it documents that several established cross-linking search engines report 1% FDR while delivering between 2.2% and 5.7% actual FDR, and offers a pipeline whose estimates stay calibrated. Weights are distributed in Triton-compatible form through Zenodo and served publicly on Koina; the training corpus is public, and the accompanying preprint carries a CC-BY-ND license.